A comparison of mixture and vector space techniques for translation model adaptation

A Comparison of Mixture and Vector Space Techniques for Translation Model Adaptation Boxing Chen Boxing.Chen@nrc-cnrc.gc.ca Roland Kuhn Roland.Kuhn@nrc-cnrc.gc.ca George Foster George.Foster@nrc-cnrc.gc.ca National Research Council Canada, Ottawa, Canada Abstract In this paper, we propose two extensions to the vector space model (VSM) adaptation tech- nique (Chen et al., 2013b) for statistical machine translation (SMT), both of which result in significant improvements. We also systematically compare the VSM techniques to three mix- ture model adaptation techniques: linear mixture, log-linear mixture (Foster and Kuhn, 2007), and provenance features (Chiang et al., 2011). Experiments on NIST Chinese-to-English and Arabic-to-English tasks show that all methods achieve significant improvement over a com- petitive non-adaptive baseline. Except for the original VSM adaptation method, all methods yield improvements in the +1.7-2.0 BLEU range. Combining them gives further significant improvements of up to +2.6-3.3 BLEU over the baseline. 1 Introduction The translation of a source-language expression to a target language might differ across genres, topics, national origins, and dialects, or the author’s or publication’s style. The word “domain” is often used to indicate a particular combination of all these factors (Chen et al., 2013b). Sta- tistical machine translation (SMT) systems are trained on bilingual parallel and monolingual data. The training data vary across domains, and translations across domains are unreliable. Therefore, we can often get better performance by adapting the SMT system to the test domain. Domain adaptation (DA) techniques for SMT systems have been widely studied. Ap- proaches that have been tried for SMT model adaptation include self-training, data selection, data weighting, context-based DA, and topic-based DA. etc. We will review these techniques in the next Section. Among all these approaches, data weighting has received most attention, it assigns each data item a weight according to its closeness to the in-domain data. Mixture model and vector space model adaptation (Chen et al., 2013b) are two data weighting tech- niques. Both of them assume the existence of N bilingual sub-corpora as training data, and have weights tuned on an in-domain development corpus (dev). Mixture model adaptation assigns weights at corpus-level; sub-models trained on different domain data sets are combined linearly or log-linearly (Foster and Kuhn, 2007). Provenance features (Chiang et al., 2011) compute a separate set of lexical weights for each sub-corpus, and then log-linearly combined. While vector space model adaptation (Chen et al., 2013b) assigns weights at phrase-level; it weights each phrase pair in the training data with the similarity score between the phrase pair and the in-domain dev data. In this paper, we first propose two extensions to the original vector space model adaptation, thereby invent two new data weighting adaptation methods based on the vector space model. The first one is grouped VSM adaptation, which classifies the in-domain data to several groups, Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 124 such as general-domain, domain-specific phrase pair, then assigns several weights to the phrase pair with the similarity scores between the phrase pair and the dev data …

Content truncated to 500 words